Worldmodeldata bets on video game data to train world models

British startup Worldmodeldata wants to organize controller inputs, images, and other data collected from video games to train world models, a class of systems designed to understand physics and act in the physical environment. The company, advised by Yann LeCun, acts as an intermediary between AI labs and game studios, which already produce large volumes of varied data about actions and consequences. Worldmodeldata says it has licensed nearly 1 million hours of data from studios behind popular games, whose names were not disclosed, and plans eventually to create ways to compensate individual players.

The hypothesis that more video game data will improve these models has not yet been fully tested. Researchers cited by Wired say the material could help capture unusual situations that are difficult to obtain through manual demonstrations with people, robots, and sensors, while the startup believes game data will eventually make up most of the training material, supplemented by information specific to each real-world environment.